Add Phase 3 & 4 resilience tests: chaos, data integrity, and concurrency (tests 58–63)

[Copilot]

Continues executing the resilience testing plan in docs/resilience-testing.md. Phases 1, 2, and 5 were already complete (tests 38–57). This PR adds Phases 3 and 4.

New tests (58–63)

Test	Plan	What it validates
58_kill_worker_mid_execution	D1	pg_terminate_backend on the BGW → PG auto-restarts it in ≤5s (epoch sentinel change confirms); in-flight instance must reach a terminal state
59_stuck_instances	E2/E3	Signal-waiting instance stays running indefinitely — no idle timeout; df.cancel() is the only exit
60_orphaned_nodes	E1	Deleting a df.instances row leaves df.nodes rows intact (no FK cascade); df.status() returns gracefully
61_table_bloat	E4/E5	Instance and node row counts grow proportionally after 10 runs; confirms zero automatic GC across df/duroxide schemas
62_concurrent_sessions	F1	10 independent dblink connections each call df.start(); all produce distinct IDs and complete
63_shared_variable_race	F4	Variable snapshot-at-start semantics verified; cross-session last-writer-wins overwrite on df.vars demonstrated and documented

New key findings

ID	Finding	Severity
F8	No FK between df.instances / df.nodes — orphaned nodes accumulate silently	Design gap
F9	No GC for completed instances or duroxide history — unbounded table growth	Design gap
F10	df.vars is per-owner-global — concurrent sessions race on the same key (last writer wins)	Design gap
F11	No idle timeout for signal-waiting instances — df.cancel() is the only escape	Design gap

Test runner

Tests 58, 60, 62, 63 added to the superuser (postgres) override list in scripts/test-e2e-local.sh — they require pg_terminate_backend, direct RLS-bypassing table access, or dblink with postgres credentials.

Remaining gaps (not addressed here)

D2 (PostgreSQL crash recovery), F2/F3 (signal/cancel vs. completion races), and F5 (concurrent df.status() lock contention) require a shell-level or timing-sensitive test harness beyond what SQL-only E2E tests support.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.